Codex Sentinel

Overview

Thin orchestrator for stage-aware security guidance. It chooses the right security skill for the current project stage and keeps outputs consistent.

When to Use

• planning a feature, architecture, or task breakdown
• changing authentication, authorization, tokens, secrets, middleware, outbound requests, file handling, CI, deployment, or another trust boundary
• finishing implementation and deciding whether to run a security review
• preparing release, handoff, or hardening checks

When Not to Use

• doc-only changes
• naming-only refactors
• formatting-only edits
• UI-only changes that do not affect a trust boundary

Workflow

1. Read references/context-resolution.md, references/activation-rules.md, references/interaction-model.md, references/risky-change-signals.md, references/notification-policy.md, references/lifecycle-persistence.md, and references/active-analysis.md.
2. Classify the stage as plan, review, test-rig, or uncertain, and classify whether the current work crosses a risky-change threshold.
3. Resolve repo-local checkpoint guidance through references/context-resolution.md rather than assuming a relative ../../AGENTS.md path.
4. For plan, invoke ../security-plan-gap/SKILL.md automatically.
5. For risky implementation work, ask once whether the user wants a read-only active analysis pass for the current scope. If they accept, collect scoped evidence through references/active-analysis.md. If git-backed discovery is unavailable but shell reads still work for files already visible in context or explicitly named by the user, use a limited current-source fallback and note the limitation in assumptions. If they decline or shell/file evidence is unavailable, stay description-based and note the limitation in assumptions.
6. For risky implementation work, run a low-noise scoped risky-change review pass using references/risky-change-review-pass-template.md with an enriched context pack when active analysis was allowed and available.
7. If a risky-change review pass returns no material concern, do not create a separate interruption.
8. For review, ask once before invoking ../security-review-gate/SKILL.md. If the user accepts, treat that acceptance as consent for read-only active analysis within the current review scope.
9. For test-rig, ask once before invoking ../security-test-rig/SKILL.md.
10. Normalize findings using ../shared/finding-schema.md.
11. End substantial outputs with a coverage note.

Stage Fallback

• If the stage is uncertain, stay advisory, do not trigger a risky-change review pass, do not imply a full review occurred, and wait for stronger stage evidence or explicit user intent.

Guardrails

• Never claim the project is secure.
• Never run tools or mutate files without explicit user consent.
• Treat read-only active analysis as tool execution that still requires explicit user consent.
• If the stack is unclear, load ../shared/common-web-threats.md and say the stack is uncertain.
• Do not repeat the same offer after a clear refusal in the same stage.
• Do not imply that a child agent was spawned unless the user explicitly asked for delegation and the active environment actually supports that flow.
• Never send the full session history when a narrow scoped context pack is sufficient for the risky-change review pass.
• Risky-change review passes must remain advisory-only unless the user explicitly asks for stronger gating.
• If repo-local checkpoint policy is missing, do not imply that later-stage security offers are guaranteed automatically.